Teacher-Student Semi-supervised Strategy for Abdominal CT Organ Segmentation

Semi-supervised abdominal multi-organ segmentation is a challenging topic. In recent years, many methods for automatic segmentation based on fully supervised deep learning have been proposed. However, it is very expensive and time-consuming for experienced medical practitioners to annotate a large number of pixels. Therefore, more researchers focus on semi-supervised learning in abdominal organ and tumor segmentation. In this paper, we adopt a classical Teacher-student semi-supervised strategy to perform the task of abdominal organs and tumor segmentation. Unet is used as the architecture for the segmentation network. Based on the Unet network structure, we add the Inception block and SEBlock to achieve more accurate segmentation. Inception block is its ability to simultaneously capture features at multiple different scales. By introducing SEBlock, the model can better focus on specific information relevant to the task while reducing attention to noise or irrelevant information. Besides, we combine Cross Entropy Loss and Dice Loss as loss functions to improve the performance of our method. We apply a teacher-student model with exponential moving average (EMA) strategy to update the network model parameters. The organs and tumor mean DSC on the public validation set was 85.39%, 18.30% respectively, the organs and tumor mean NSD was 89.36%, 6.44% respectively. And the average running time and the area under GPU memory-time curve 35.54 s, 38175.35.